"Best" machine for xtal computing?

Geoff Barton gjb at bioch.ox.ac.uk
Fri Feb 7 13:47:02 EST 1997


Kay Diederichs wrote:
> 
> Yoram Puius (puius at aecom.yu.edu) wrote (edited):
> :
> : We are looking to buy a number-cruncher for protein crystallography
> : (i.e. very fast in running X-PLOR, TNT, XDS, etc.)  We have mostly
> : SGIs as combination graphics/processing stations, but we're looking
> : into a number cruncher described by SGI as an
> :
> : "Origin Server, 4XR1000, 180MHz, 1MB cache, 128MBmem, 2X2.2GB Sys Disk
> : (2 Twr no skins) IRIX for 4CPUs"
> :
> : I am assuming this means a machine with four R1000 CPUs, which are
> : certainly faster than what we have around, although I don't know
> : if X-PLOR has been optimized for the R10000 machines yet.  I am
> : not counting on anyone out there parallelizing existing software,
> : by the way.
> 
> The Origin server family replaces the Challenge server family.
> You might be able to buy Challenge/PowerChallenge still, but the price/
> performance ratio is (much) better for the Origin family. They are very
> expandable, up to supercomputer range.
> Do not expect X-PLOR (current version) to run significantly faster
> on SGI multiprocessor, as it doesn't parallelize well. Of course you can
> run several X-PLOR jobs at the same time on a multiprocessor machine, and
> that is probably the preferred mode, if your problem is tractable by that.
> I/O is not an issue with X-PLOR, and probably not even with XDS.

You may also want to look at the 4Mbyte cache size for the R10000.  The extra
cache can make a difference to speed of some codes, particularly floating point.

You CERTAINLY want more than 128MBytes of RAM on a 4 processor machine.  Go for
at least 512MBytes.  Make sure that your quote includes the parallel 
compilers etc.

Best to benchmark your code on all available machines and certainly look at 
DEC too.  Don't believe SPECmark figures, except within a manufacturers own range.

DEC's cpus are definitely competitive (probably faster) 
than the R10,000 at the moment, though I do not know what the cost penalty is.
If you don't want to go to more than 12 CPUs, then DEC may be a better choice, but
SGI may be flexible on price if they think they are going to lose the deal to DEC.

Of course, speed is not everything.  If your lab is full of SGI's at the moment,
then the extra work involved in having two different computer types may outweigh
any speed/cost advantage.

Hope this helps,

Geoff.
------------------
Geoffrey J. Barton, Laboratory of Molecular Biophysics, University of Oxford, 
Rex Richards Building, South Parks Road, Oxford OX1 3QU, U.K.
mailto:geoff.barton at ox.ac.uk, Tel: +44 1865 275368, Fax: +44 1865 510454, 
ftp://geoff.biop.ox.ac.uk, http://geoff.biop.ox.ac.uk




More information about the Xtal-log mailing list